use SHA512 when seeding MersenneTwister

[rfourquet]

This hashes the provided seeds before feeding dSFMT initialization routine.
The passed seeds are still stored as-is in the MersenneTwister.seed field.

Goals:

• possibly lead to more independance between two streams corresponding to two
  user-provided seeds
• by feeding VERSION.minor and VERSION.major to the hashing process, this
  guarantees that random streams change between minor releases, helping users
  to not believe in stream stability

---

Closes #37165.

Performance

Random.seed!(rng, 0) takes, on my machine,

• 12.655 μs on master
• 13.460 μs on this branch but using SHA1 (6% slowdown)
• 14.320 μs on this branch, with SHA2_512 (13% slowdown)

As seeding is usually not a time-critical operation, the slowdown here seems acceptable (it will probably be a different story when Xoshiro lands as the default RNG...)

Handling our tests

Many of our tests use Random.seed!(), and these would have to be changed for all minor releases. Alternatives include:

• having an escape hatch, e.g. seed!(seed, hash=false), or Random.seednohash(seed)
• implement a simple RNG with stable streams in our tests...

[StefanKarpinski]

Is this needed?

[rfourquet]

As SHA is accessed, I guess yes? I prefer using qualifed names like SHA.update! here. Or did I misunderstand the question?

[StefanKarpinski]

Ok, I see: it's used in RNG.jl but not in this file. Might be clearer to move the import SHA into the file that uses it, but this is valid as well.

[rfourquet]

Oh ok I understand. it's just that Random.jl already happens to have all the import/using directives necessary for the other files it includes.

[StefanKarpinski]

Aside from my question (and CI), I approve very much.

[rfourquet]

and CI

Yes, I'm not sure what to do here. We don't want to update the seeds in tests for each minor release (finding a working seed is very tedious). I think I will go with a the hack seed!(rng, hash=false) for now.

[rfourquet]

Also, many jldoctests will stop working, and there we don't want to expose seeding with hash=false. I would just remove the jldoctest part, but wanted to gather suggestions before going this path. The obvious alternative is to update the examples for each release, which would be less annoying than updating all the tests (in LinearAlgebra in particular), but still tedious.

[StefanKarpinski]

If it's painful for us, then perhaps it's too painful to inflict on everyone else too. Finding seeds that work shouldn't really be that hard: if the tests are valid, then most seeds should work. The trouble, I guess, is that while most seeds should work for each test, the chance that a given global seed will allow all tests to pass is slim. It seems like each stochastic test should get its own seed then and you just try new random seeds for any test that fails and most choices should work.

[rfourquet]

For doctests, I don't have a strong opinion what should be done. But for tests, I think it's easier than in the past to find fitting seeds in Base, probably because indeed seeding is done at a more granular level. I don't mind at all doing it for this PR, but the simple PR of bumping the minor version of Julia will likely not pass tests.

If it's painful for us, then perhaps it's too painful to inflict on everyone else too

So for tests, that's kinda the point of the PR, to motivate people to not rely on MersenneTwister for this kind of tests. But in Base we don't really have an alternative.

[rfourquet]

If this is to be done, consider putting the hashing logic within make_seed, cf. #41558 (comment)